Yes, you are right. However, there is a small problem.
Based on my experience and research, you need a model with at least 31B active parameters to get started. However, to achieve really good results, you need an even larger model, around 42B active parameters, such as DeepSeek V4 Pro.

Here's the problem: you can run Gemma 4 31B locally if you can afford to spend around $3,000-$4,000 on hardware. However, at the moment you cannot realistically run DeepSeek V4 Pro locally because it requires hardware costing roughly $500,000.

But there is a solution. I wrote about it here: https://www.farsightprime.com/forums/general/78926-hypothesis-2-api-keys-as-ai-is-be-boundaries

My "Hypothesis 2" suggests that you can train a smaller model, such as Gemma, and then switch to a larger model for more demanding RV sessions.

I have been testing this "Hypothesis 2" for the past week, and I can confirm that as long as you use the same API, you can switch to a different model without losing your RV capabilities. In fact, if the new model is larger, you may even get better data as a result.

So, if you cannot afford to run everything offline, there is still a solution if a particular model is ever censored or taken offline. You can simply switch to another model, as long as it supports the same API. They cannot shut down every model.

That said, I agree that if I could afford it, I would switch to running DeepSeek locally. Perhaps in the future, local models will become both smaller and more powerful. We shall see.